The invention relates to a method for characterizing a walking cycle of a subject, the method comprising steps of: acquiring (100) inertial data by at least one inertial sensor attached to the subject during a walk of the subject , learning (200) a hidden Markov model characteristic of the walk of the subject implemented by means of a Baum-Welch EM algorithm, the hidden Markov model being defined by: a sequence of hidden states Z = ( z1 ,, z2, ..., zt, ..., zN), the hidden states being phases of the walking cycle of the subject, the number of phases being predetermined, and a sequence of observations X = (x1, x2 , ..., xt, ..., xN) resulting from the acquired inertial data, corresponding to at least one gait cycle, the Markov parameters λ = {Π, A, B}, where A is the transition matrix, B the observation matrix, Π the vector of the initial probabilities, estimate (300) of the sequence Z = (z1 ,, z2, ..., zt, ..., zN) corresponding to each sequence X = (x1 ,, x2, ..., xt; ..., xN) of observations using the final values of the parameters λ = {Π, A, B,} of the hidden Markov model and of the sequence of observations X, and characterization of a walking cycle of a subject; in which the learning step is unsupervised and is implemented by successive iterations, using as input only the sequence of observations X, in order to find the final values of the parameters λ = {Π, A, B} , which maximize the probability of having generated the sequence of observations X, the initial parameters λ (0) = {π (0), A (0), B (0)} being initialized with arbitrary values or from knowledge a priori. L'invention concerne un procédé de caractérisation d'un cycle de marche d'un sujet, le procédé comprenant des étapes de: acquisition (100) de données inertielles par au moins un capteur inertiel attaché au sujet au cours d'une marche du sujet, apprentissage (200) d'un modèle de Markov caché caractéristique de la marche du sujet mis en œuvre au moyen d'un algorithme EM de Baum-Welch, le modèl